Graph Neural Network‐Enabled Robust AIS/Radar Data Association for Intelligent Monitoring of Ship Safety

ABSTRACT In intelligent waterborne transportation systems, wide‐range perception devices such as AIS receivers and navigation radars are commonly deployed to enhance the monitoring capability of ship traffic. The AIS data provides ship identity and dynamic navigational information. The radar complementarily produces higher frequency position and movement data, but it has detection blind spots. To further improve ship intelligent surveillance, it becomes necessary to associate the AIS and radar data to simultaneously capture identity, high‐frequency and dynamic information for the ships of interest. However, AIS and radar data suffer from time asynchrony, spatial differences, and environmental noise. The conventional track‐based AIS/radar association methods heavily rely on precise tracking results and stable trajectories, and also suffer from high computational complexity. We propose a graph neural network (GNN)‐enabled robust AIS/radar data association method. The association task is formulated as a graph matching problem between imbalanced bipartite graphs. We then construct a paired graph dataset from AIS and radar data. A specialized GNN architecture is developed to extract and aggregate spatial distribution features of ship targets and their neighboring nodes. Contrastive learning techniques are employed to formulate the loss function for feature extraction networks. Subsequently, cosine similarity metrics between extracted ship distribution features are computed to construct a cost matrix for target association. The final linear assignment problem is resolved using the Hungarian algorithm, enabling precise AIS/radar data association. This methodology demonstrates significant improvements in both computational efficiency and matching accuracy.